From the Image Tokenization talk from Google and Verisilicon for Edge Vision Summit

Using VAE for tokenization (?????)

Tokenizer is a Feature Extractor

Tokenization Creates a form of data compression

Neural Cascade

why isn’t there a bfloat8